If you’ve tried to generate or edit an image in ChatGPT recently, you might have run into a brick wall. The progress bar starts, the model acts like it’s processing, and then it hangs indefinitely on a single, frustrating message: “One last tweak…” or “Finishing touches.”
And if you refresh the page or try to back out, it doesn’t clear the bottleneck. The web UI simply restarts the image generation process from the beginning of that step, only to hit the exact same server wall and get stuck in the loop all over again.
This isn’t a problem with your browser, your internet connection, or your local DNS cache. It is a known, widespread infrastructure glitch buried deep in OpenAI‘s tool-calling pipeline – and there are two very specific triggers causing it.
The Triggers: Live Web Links and Massive Prompts
While the hang can happen randomly, users have isolated two massive, consistent triggers that reliably break the image generator: asking ChatGPT to check a live link before creating an image, and submitting highly detailed, text-heavy prompts.
Here is what is happening under the hood in both scenarios:
- The Live Link Handover Failure: When you ask the AI to analyze a live URL (like a webpage or design reference) and translate it into a visual asset, the backend has to juggle two distinct internal tools: the web browser tool and the DALL-E 3 image generation engine. The browser successfully scrapes the webpage data, but when it attempts to pass that massive contextual payload over to the image generation endpoint, the pipeline stalls.
- The Long Prompt Payload Bloat: Even without a link, writing an extremely detailed prompt with dozens of micro-instructions can cause the exact same issue. When the text description is too long, or when you are deep in a chat thread with a massive history of back-and-forth edits, the background data payload bloats. The image engine struggles to process the massive token context alongside the image generation command, causing the backend connection to time out.
In both cases, the text interface thinks the job is almost done (hence “One last tweak…”), while the rendering engine is waiting in total silence, never receiving the full completion signal.
Eventually, if you are patient enough to wait, you will get the image generation failed message.
How to Bypass the “One Last Tweak” Loop
OpenAI hasn’t pushed a permanent patch downstream to fix this routing failure yet, but you don’t have to wait for them to clean up their infrastructure. Here is how you can unstick your account and protect your workflow:
1. Split the Workflow (The Link Fix)
If you need an image based on a webpage, do not combine the requests.
- Step 1: Paste the link and say: “Analyze this page and summarize the visual style, colors, and key elements in plain text.”
- Step 2: Copy that text output, start a brand new chat, and paste that raw text summary as your prompt for the image. By stripping out the live browsing tool entirely, the image engine handles the request without a hitch.
You could also do this in one step – just copy/paste the content from your link in the prompt and ask for an image to go with it.
2. The Cross-Platform Force-Close
If you are currently staring at a spinning wheel on your desktop browser, don’t press refresh just yet. (anyways refresh most likely won’t work, based on user feedback!)
- Open the ChatGPT mobile app on your phone and go to that exact same chat thread.
- Because the mobile app handles data streaming hooks differently, opening the thread on your phone often forces the stuck server endpoint to close, successfully rendering and displaying the image. Once it appears on your phone, refresh your desktop, and the loop will be broken.
3. Branch Your Chat to Avoid Bloat
The longer a chat thread gets, the heavier the background payload becomes. If a thread gets stuck on “One last tweak,” hover over the prompt just before the failure, click the edit/three-dots icon, and select “Branch into new chat.” This forks the conversation, ditching the corrupted execution loop while preserving all your prior context.
4. Check Your Media Library
Before you assume an image was lost to the void, check your account’s Media Gallery or Library in the sidebar. Because the backend processing sometimes completes even when the chat window crashes, the finished graphic might be sitting safely in your history, waiting to be downloaded.
In Summary
The best fix to the ChatGPT image generation issues is most likely to just simplify the prompt.
Related:
Another ChatGPT error/fix: “You’ve Reached Our Limits of Messages. Please Try Again Later.” – What the ChatGPT Error Means & How to Fix It
It’s not just ChatGPT that has the error messages: Why Does Google Gemini Keep Saying “Something went wrong (1099 or 1076)”?
Maybe it’s not so bad if ChatGPT doesn’t work for you: Could Overusing ChatGPT Harm Your Critical Thinking? A Recent Study Warns
If you love using ChatGPT, you’re not alone: 9 in 10 Gen Zers Can’t Imagine Life Without ChatGPT
ChatGPT can help promote your website: How to Check if Your Website is Showing Up in ChatGPT, Perplexity, Gemini, or Other AI Answers
ChatGPT apparently can even help you win the Powerball: How a Michigan Woman Used ChatGPT to Win $100,000 in the Powerball
But using ChatGPT can also cause you to lose out on landing that new job: Anne Hathaway Saw Something Strange in Job Applications Using AI
And don’t forget how AI can and does get it wrong: Google AI Says to put Elmer’s Glue in Your Pizza Sauce…How Smart Is AI Really?
Ok – now check your ChatGPT image generation – did you get the Fail message yet? Follow the recommended fix!
Looking for a website builder that is not AI but lets you add some AI generated code to it when you want? Learn more about UltimateWB! We also offer web design packages if you would like your website designed and built for you.
Got a techy/website question? Whether it’s about UltimateWB or another website builder, web hosting, or other aspects of websites, just send in your question in the “Ask David!” form. We will email you when the answer is posted on the UltimateWB “Ask David!” section.